p1atdev / dart-v1-sft

Dart (Danbooru Tags Transformer) v1

This model is a fine-tuned Dart ( Da nboo r u T ags Transformer) model that generates danbooru tags.

Demo: 🤗 Space

If you are a developer and want to finetune, it's recommended using the base version, p1atdev/dart-v1-base , instead

Usage

Using AutoModel

🤗 Transformers library is required.

pip install -U transformers

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

MODEL_NAME = "p1atdev/dart-v1-sft"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True) # trust_remote_code is required for tokenizer
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.bfloat16)

prompt = "<|bos|><rating>rating:sfw, rating:general</rating><copyright>original</copyright><character></character><general><|long|>1girl<|input_end|>"
inputs = tokenizer(prompt, return_tensors="pt").input_ids

with torch.no_grad():
  outputs = model.generate(inputs, generation_config=model.generation_config)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# rating:sfw, rating:general, 1girl, ahoge, braid, closed eyes, collared dress, dress, flower, full body, hair flower, hair ornament, long hair, night, night sky, outdoors, parted lips, pink flower, pink hair, short sleeves, sky, solo, straight hair, sunflower, very long hair, white flower

You can use tokenizer.apply_chat_template to simplify constructiing of prompts:

inputs = tokenizer.apply_chat_template({
  "rating": "rating:sfw, rating:general",
  "copyright": "original",
  "character": "",
  "general": "1girl",
  "length": "<|long|>"
}, return_tensors="pt", tokenize=True) # tokenize=False to preview prompt
# same as input_ids of "<|bos|><rating>rating:sfw, rating:general</rating><copyright>original</copyright><character></character><general><|long|>1girl<|input_end|>"
with torch.no_grad():
  outputs = model.generate(inputs, generation_config=generation_config)

See chat_templating document for more detail about apply_chat_template .

Flash attention (optional)

Using flash attention can optimize computations, but it is currently only compatible with Linux.

pip install flash_attn

Accelerate with ORTModel

🤗 Optimum library is also compatible, for the high performance inference using ONNX.

pip install "optimum[onnxruntime]"

Two ONNX models are provided:

• Normal
• Quantized

Both can be utilized based on the following code:

import torch
from transformers import AutoTokenizer, GenerationConfig
from optimum.onnxruntime import ORTModelForCausalLM

MODEL_NAME = "p1atdev/dart-v1-sft"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)

# normal version
ort_model = ORTModelForCausalLM.from_pretrained(MODEL_NAME)

# qunatized version
# ort_model = ORTModelForCausalLM.from_pretrained(MODEL_NAME, file_name="model_quantized.onnx")

inputs = tokenizer.apply_chat_template({
  "rating": "rating:sfw, rating:general",
  "copyright": "original",
  "character": "",
  "general": "1girl",
  "length": "<|long|>"
}, return_tensors="pt", tokenize=True)

with torch.no_grad():
  outputs = ort_model.generate(inputs, generation_config=model.generation_config)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Prompt guide

Due to training with a specialized prompt format, natural language is not supported .

The trained sentences are essentially composed of the following elements, arranged in the strict order shown below:

• <|bos|> : The bos (begin of sentence) token

• <rating>[RATING_PARENT], [RATING_CHILD]</rating> : The block of rating tags

  • [RATING_PARENT]: rating:sfw , rating:nsfw
  • [RATING_CHILD]:
    • if [RATING_PARENT] is rating:sfw : rating:general , rating:sensitive
    • else: rating:questionable , rating:explicit

• <copyright>[COPYRIGHT, ...]</copyright> : The block of copyright tags.

  • [COPYRIGHT, ...]: All supported copyright tags can be seen in here

• <character>[CHARACTER, ...]</character> : The block of character tags.

  • [CHARACTER, ...]: All supported character tags can be seen in here

• <general>[LENGTH_TOKEN][GENERAL, ...]<|input_end|>[COMPLETION]</general> : The block of general tags.

  • [LENGTH_TOKEN]: A token to specify total amount of general tags.
    • Avaiable:
      • <|very_short|> : less than 10 tags
      • <|short|> : less than 20 tags
      • <|long|> : less than 40 tags (recommended)
      • <|very_long|> : more than 40 tags
  • [GENERAL, ...]: All supported general tags can be seen in here
  • <|input_end|> : A tag to show the end of input. Set this token at last of prompt.
  • [COMPLETION]: The model complete tags in alphabetical order.

• <|eos|> : The eos (end of sentence) token

• Tags other than special tokens are separated by commas.

• You can place tags in any order you like in each block.

Example sentence:

<|bos|><rating>rating:sfw, rating:general</rating><copyright>vocaloid</copyright><character>hatsune miku</character><general><|long|>solo, 1girl, very long hair<|input_end|>blue hair, cowboy shot, ...</general><|eos|>

Therefore, to complete the tags, the input prompt should be as follows:

1. without any copyright and character tags

<|bos|><rating>rating:sfw, rating:general</rating><copyright></copyright><character></character><general><|very_long|>1girl, solo, cat ears<|input_end|>

1. specifing copyright and character tags

<|bos|><rating>rating:sfw, rating:general</rating><copyright>sousou no frieren</copyright><character>frieren</character><general><|long|>1girl, solo, from side<|input_end|>

Model Details

Model Description

• Developed by: Plat

• Model type: Causal language model

• Language(s) (NLP): Danbooru tags

• License: Apache-2.0

• Demo: Avaiable on 🤗Space

Bias, Risks, and Limitations

Since this model is a pre-trained model, it cannot accommodate flexible specifications.

Training Details

Training Data

This model was trained with:

• isek-ai/danbooru-tags-2023 : 6M size of danbooru tags dataset since 2005 to 2023

Only data from 2020 onwards was used for SFT.

Training Procedure

Trained using 🤗 transformers' trainer.

Preprocessing

Preprocessing was conducted through the following process:

1. Remove data where general tags is null.
2. Remove general tags that appear less than 100 times.
3. Remove undesirable tags such as watermark and bad anatomy .
4. Remove based on the number of tags attached to a single post (following rules):

• Remove if more than 100 for general tags.
• Remove if more than 5 for copyright tags.
• Remove if more than 10 for character tags.

1. Remove posts created before 2020
2. Set length token according to each tags length
3. Shuffle some tags in the following rule:

• Include people tags (e.g. 1girl , no humans ) tags in the shuffle-group with a 95% probability, and do not do so with a 5% probability.
• Get tags at a random percentage between 0% and 75% to create a shuffle-group.
• Shuffle tags in shuffle-group and concatnate with <|input_end|> token and remains in alphabetical order.

1. Concatnate all categories

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 1

Evaluation

Evaluation has not been done yet and it needs to evaluate.

Technical Specifications

Model Architecture and Objective

The architecture of this model is OPT (Open Pretrained Transformer) , but the position embeddings was not trained.

Compute Infrastructure

In house

Hardware

1x RTX 3070 Ti

Software

• Dataset processing: 🤗 Datasets
• Training: 🤗 Transformers
• Optimizing: 🤗 Optimum
• SFT: 🤗 TRL

More Information [optional]

[More Information Needed]